NVIDIA's $2 billion deal with Groq yields results as Groq 3 LPX chips enter mass production

Breaking News from Dòngchá Beating AI: NVIDIA’s Groq 3 LPX has entered full-scale mass production. In December last year, NVIDIA acquired the licensing rights to Groq’s technology for approximately $20 billion and recruited its core team, including founder Jonathan Ross. Eight months later, the first hardware product from this deal has officially launched. The Groq 3 LPX is a system specifically designed to accelerate large model inference, incorporating 256 Groq 3 LPU chips. In tests by Artificial Analysis, when processing a Gemma 4 31B model with an input of 100,000 tokens, the Groq 3 LPX achieved a throughput of 3,431 output tokens per second. By comparison, the fastest publicly available API on the market at the time delivered only around 870 tokens per second—meaning the Groq 3 LPX is nearly four times faster. In real-world deployments, NVIDIA’s Rubin GPUs handle heavier computational tasks, while the Groq 3 LPX is dedicated to accelerating token generation. For coding agents that repeatedly call models dozens or hundreds of times in succession, this significantly reduces waiting time for model responses. Nebius will be the first AI cloud provider to adopt the Groq 3 LPX. Interestingly, Groq itself is also among the first adopters and will deploy the system alongside Dell. Last year, Groq licensed its technology to NVIDIA; this year, it is using NVIDIA’s version of “Groq.”